Two AI labs looked at the same problem and reached opposite conclusions.
OpenAI is previewing a system called Private Safety Processing to select customers. It's designed to catch misuse that spans multiple conversations without retaining the underlying data at all.
Anthropic went the other direction. Its most capable models now require 30 days of data retention specifically so the company can review flagged activity after the fact.
Both companies frame this as a response to the same underlying threat: attacks and misuse patterns that only become visible when you can see multiple interactions together, not just one request in isolation.
Anthropic's Bet: Retain to Detect
The policy applies to what Anthropic calls Covered Models, currently Fable 5 and Mythos 5, and any future model with similar capabilities. Prompts and outputs are held for 30 days, then automatically deleted unless flagged for a safety review or required by law.
Access is tightly controlled. By default no Anthropic employee can read retained conversations, and human review only happens through a logged, limited-reviewer process when automated systems flag something.
Anthropic hasn't pretended this is a costless decision. In a risk report, the company described the policy as one it expects will be unpopular with customers used to zero retention, and acknowledged it "pose[s] real risks to our business success, especially if competitors do not follow."
That's an unusually direct admission for a company making a product decision.
Consumer-facing plans aren't affected either way. Anthropic's retention policy applies specifically to Covered Models accessed through its API and platform partners, not to its standard consumer subscriptions, which already operated under a retention period before this policy existed.
OpenAI's Bet: Detect Without Retaining
OpenAI's approach tries to solve the same problem differently. Private Safety Processing analyzes patterns across related conversations, looking for things like a bad actor spreading malware-building requests across multiple sessions to avoid detection in any single one.
The system is built to send OpenAI a narrow safety signal rather than the underlying prompts or responses. Customer data can stay on the customer's own infrastructure, or be stored with encryption keys the customer controls.
OpenAI plans a broader rollout and a technical white paper in September.
Both Companies Admit This Is a Tradeoff
Neither company is claiming to have solved the underlying problem cleanly.
OpenAI acknowledges it still needs some visibility over time to catch multi-session abuse, even without full data access. Anthropic acknowledges its retention window creates exactly the privacy exposure some enterprise customers have specifically tried to avoid.
The split comes as both labs face pressure on a related front.
OpenAI recently paused parts of its own model training over safety concerns following the Hugging Face incident MAIN has covered. Anthropic has said it doesn't currently see the same need.
What This Means for Miami
Miami businesses evaluating either company's models for sensitive workloads, legal research, health data, financial records, now have a genuinely different tradeoff to weigh depending on which vendor and which specific model they're using.
A firm that needs zero data retention as a hard compliance requirement may find OpenAI's approach easier to work with once Private Safety Processing rolls out broadly. A firm more focused on catching sophisticated, multi-session misuse attempts might view Anthropic's retention window as the safer bet, provided the access controls hold up as described.
Neither approach is obviously correct. The right one depends on what a given Miami business is actually more afraid of: data exposure, or missed detection.
That question is worth asking explicitly during vendor selection rather than assuming either company's default settings match a firm's actual risk tolerance.
Legal and compliance teams evaluating AI vendors this year should treat data retention policy as a line item worth negotiating, not just a checkbox to confirm exists.